Lichtle, Jonathan, Meddeb, Aymen, Birot, Antonin, Niang, Ibrahima, Gastaud, Lisa, Hanafi, Riyad, Dujardin, Sofiane, Selomme, Corentin, Cachard, Guillaume, Hamroun, Aghiles, Kuchcinski, Grégory, Kim, Su Hwan, Guellec, Bastien Le · Nov 17, 2026
Study finds vision-language model agents often kept an early, wrong diagnosis on emergency scans
Markdown · JSON · Original source
The paper’s authors evaluated six vision-language model (VLM) agents reading 100 emergency CT and MRI examinations from a public archive. Across four ways of presenting or navigating images, pooled accuracy ranged from 32.6% with free navigation to 42.5% with key slices in one request. On 30 cases also read by radiologists, freely navigating models were correct in 30.0% of runs, compared with 77.0% for radiologists. The authors also found that models often formed an early hypothesis and kept it: after a wrong first hypothesis, final answers were correct in 12.4% of runs. Median stated confidence was 90%, while 38.9% of final answers were correct. The paper argues that both navigation behavior and early commitment matter: letting an agent inspect images freely did not make it more accurate in this study, and it seldom corrected a wrong initial hypothesis. These results concern the study’s retrospective set of examinations and tested conditions; they do not establish how the agents would perform in clinical practice.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 17, 2026 (2026-11-17).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-19T08:48:10.688Z. First observed: 2026-11-19T05:01:24.366Z. Retrieved: 2026-11-19T08:47:07.649Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The study was retrospective and used 100 examinations from a public archive, rather than establishing performance across emergency imaging in clinical practice. Nine radiologists read only 30 of the 100 cases, so the reported comparison with radiologists covers that subset. Answers were graded by a judge validated against two radiologists; the abstract does not provide further details about the validation.
Open question
Would the same accuracy and early-hypothesis patterns appear in a broader set of emergency CT and MRI examinations?
Source-stated
The authors evaluated six vision-language model agents on 100 emergency CT and MRI examinations from a public archive, using four exploration conditions.
six VLMs read 100 emergency CT and MRI examinations published in a public archive
Source-stated
Pooled accuracy was lowest with free navigation and highest when key slices were presented in one request.
Pooled accuracy was lowest under free navigation (32.6%; 391/1200), followed by navigation from a key slice (37.9%; 455/1200; P = .003), and highest with all slices (42.2%; 506/1200) or key slices (42.5%; 510/1200) in one request
Source-stated
On the 30 cases read by radiologists, freely navigating models were correct in fewer runs than radiologists.
freely navigating models were correct in 30.0% of runs (108/360) versus 77.0% for radiologists (208/270; P
Source-stated
The authors report that models often retained an early hypothesis, and that final answers were rarely correct after a wrong first hypothesis.
They adopted an early hypothesis and usually kept it, seldom correcting a wrong one.
Source-stated
The authors report a gap between median stated confidence and the share of final answers that were correct.
Median stated confidence in the final answers was 90%, whereas 38.9% of them were correct
Escape Meridian interpretation
Because the evaluation used a retrospective set of 100 examinations and specific tested conditions, its results do not by themselves establish performance in clinical practice.
In this retrospective study, six VLMs read 100 emergency CT and MRI examinations
Source and contributor context
Lichtle, Jonathan, Meddeb, Aymen, Birot, Antonin, Niang, Ibrahima, Gastaud, Lisa, Hanafi, Riyad, Dujardin, Sofiane, Selomme, Corentin, Cachard, Guillaume, Hamroun, Aghiles, Kuchcinski, Grégory, Kim, Su Hwan, Guellec, Bastien Le · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 7fed72cb-a2fc-48ef-8c6a-1f740238e622.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-19T08:48:10.688Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 169fe74f71a832b475c7bfe6702526b528ca421586d78ac60e7811c7449c695c. Source text hash: 68cb0045d3555b7f78446657879077dca1d607444e4ae8d0e453e02291ef9456.
Revision and source-check history
Version 1 · 2026-11-19T08:48:10.688Z
Study finds vision-language model agents often kept an early, wrong diagnosis on emergency scans
The paper’s authors evaluated six vision-language model (VLM) agents reading 100 emergency CT and MRI examinations from a public archive. Across four ways of presenting or navigating images, pooled accuracy ranged from 32.6% with free navigation to 42.5% with key slices in one request. On 30 cases also read by radiologists, freely navigating models were correct in 30.0% of runs, compared with 77.0% for radiologists. The authors also found that models often formed an early hypothesis and kept it: after a wrong first hypothesis, final answers were correct in 12.4% of runs. Median stated confidence was 90%, while 38.9% of final answers were correct. The paper argues that both navigation behavior and early commitment matter: letting an agent inspect images freely did not make it more accurate in this study, and it seldom corrected a wrong initial hypothesis. These results concern the study’s retrospective set of examinations and tested conditions; they do not establish how the agents would perform in clinical practice.
Original source
cs.AI · Article
Source details
Original source date: Nov 17, 2026 (2026-11-17).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-19T08:48:10.688Z. First observed: 2026-11-19T05:01:24.366Z. Retrieved: 2026-11-19T08:47:07.649Z.
Limits
The study was retrospective and used 100 examinations from a public archive, rather than establishing performance across emergency imaging in clinical practice. Nine radiologists read only 30 of the 100 cases, so the reported comparison with radiologists covers that subset. Answers were graded by a judge validated against two radiologists; the abstract does not provide further details about the validation.
Open question
Would the same accuracy and early-hypothesis patterns appear in a broader set of emergency CT and MRI examinations?
Source-stated
The authors evaluated six vision-language model agents on 100 emergency CT and MRI examinations from a public archive, using four exploration conditions.
six VLMs read 100 emergency CT and MRI examinations published in a public archive
Source-stated
Pooled accuracy was lowest with free navigation and highest when key slices were presented in one request.
Pooled accuracy was lowest under free navigation (32.6%; 391/1200), followed by navigation from a key slice (37.9%; 455/1200; P = .003), and highest with all slices (42.2%; 506/1200) or key slices (42.5%; 510/1200) in one request
Source-stated
On the 30 cases read by radiologists, freely navigating models were correct in fewer runs than radiologists.
freely navigating models were correct in 30.0% of runs (108/360) versus 77.0% for radiologists (208/270; P
Source-stated
The authors report that models often retained an early hypothesis, and that final answers were rarely correct after a wrong first hypothesis.
They adopted an early hypothesis and usually kept it, seldom correcting a wrong one.
Source-stated
The authors report a gap between median stated confidence and the share of final answers that were correct.
Median stated confidence in the final answers was 90%, whereas 38.9% of them were correct
Escape Meridian interpretation
Because the evaluation used a retrospective set of 100 examinations and specific tested conditions, its results do not by themselves establish performance in clinical practice.
In this retrospective study, six VLMs read 100 emergency CT and MRI examinations
Source and contributor context
Lichtle, Jonathan, Meddeb, Aymen, Birot, Antonin, Niang, Ibrahima, Gastaud, Lisa, Hanafi, Riyad, Dujardin, Sofiane, Selomme, Corentin, Cachard, Guillaume, Hamroun, Aghiles, Kuchcinski, Grégory, Kim, Su Hwan, Guellec, Bastien Le · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 7fed72cb-a2fc-48ef-8c6a-1f740238e622.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-19T08:48:10.688Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 169fe74f71a832b475c7bfe6702526b528ca421586d78ac60e7811c7449c695c. Source text hash: 68cb0045d3555b7f78446657879077dca1d607444e4ae8d0e453e02291ef9456.
passed · 2026-11-19T08:48:10.688Z · anthropic/claude-opus-5.5
Publication: 7fed72cb-a2fc-48ef-8c6a-1f740238e622
attribution — pass: The reading attributes the study, results and conclusions to the paper's authors. It presents accuracy and confidence figures as reported findings rather than independently verified facts.
Across all exploration strategies, VLM agents remained far less accurate than radiologists.
summary — pass: The title, summary and question match the abstract. The six VLMs, 100 emergency CT/MRI exams, four conditions, the 32.6%–42.5% accuracy range, the 30.0% vs 77.0% radiologist subset comparison, early-hypothesis persistence and the 90% confidence vs 38.9% correct gap are accurate. The reading keeps the authors' core finding about early commitment. 'Key slice' is not defined, but the phrase 'ways of presenting or navigating images' gives enough context for a general reader. The question is a fair open issue.
They adopted an early hypothesis and usually kept it, seldom correcting a wrong one.
date — pass: originalPublished 2026-11-17 matches the arXiv v1 submission date, not the later feed date of 2026-11-19.
[Submitted on 17 Nov 2026]
support — pass: Each claim matches its excerpt. The clinical-practice caveat is properly labeled as interpretation and follows from the retrospective design. The limitations correctly note the 30-case radiologist subset and the LLM-judge grading. Claim 2 names key slices as highest, which matches the numbers, though all slices (42.2%) was nearly tied; this does not misstate the result. No unsupported benefits or costs are implied.
Nine radiologists read 30 of those cases. A large language model judge validated against two radiologists graded the answers.
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition reference